[Intel-gfx] [PATCH] gem_pin: Be explicit about GGTT needs

[Chris Wilson]

On Thu, Nov 07, 2013 at 10:58:24AM -0800, Ben Widawsky wrote:
> On Thu, Nov 07, 2013 at 07:54:59PM +0100, Daniel Vetter wrote:
> > On Thu, Nov 7, 2013 at 7:46 PM, Ben Widawsky <ben at bwidawsk.net> wrote:
> > > On Thu, Nov 07, 2013 at 07:36:50PM +0100, Daniel Vetter wrote:
> > >> On Thu, Nov 7, 2013 at 7:11 PM, Ben Widawsky <ben at bwidawsk.net> wrote:
> > >> >> > @@ -63,7 +65,7 @@ static void exec(int fd, uint32_t handle, uint32_t offset)
> > >> >> >     gem_exec[0].relocs_ptr = (uintptr_t) gem_reloc;
> > >> >> >     gem_exec[0].alignment = 0;
> > >> >> >     gem_exec[0].offset = 0;
> > >> >> > -   gem_exec[0].flags = 0;
> > >> >> > +   gem_exec[0].flags = LOCAL__EXEC_OBJ_NEEDS_GTT;
> > >> >>
> > >> >> This only really works with the aliasing ppgtt stuff on gen6, I'd just
> > >> >> skip the test -it's not really useful with real ppgtt.
> > >> >> -Daniel
> > >> >
> > >> > It is really useful with real ppgtt. Please rethink your assertion.
> > >>
> > >> The test links up the global gtt used by the current pin ioctl with
> > >> the ppgtt. That's not useful, except when they alias. If we want
> > >> soft-pinning, then we need a new ioctl mode to return the right
> > >> address from the right address space. Which means a new (sub)test.
> > >> -Daniel
> > >
> > > This is only indirectly related to soft pinning. No context will ever have the
> > > global GTT address space. If you want to support the pin IOCTL, (which
> > > Chris has before said he requires - the original patch series disabled
> > > it) you must do this. We already have a flag that does what we want,
> > > and, demonstrably, a test which exercises it.
> > >
> > > I do not think a new ioctl, nor a new subtest is the solution for this.
> > 
> > Ok, I'll wait until I see the code that implements this on top of
> > ppgtt. But until we have that I don't think we need to change the test
> > at all.
> > -Daniel
> 
> This is a question for Chris. If DDX only depends on the behavior with
> aliasing PPGTT, then I agree - no test change needed.

Where pin has come in handy has been to w/a hw bugs, so predicting when
and why we might need it again is difficult. As such I'd rather we
maintained its existing semantics of pinning an object in the ggtt into
the full-ppgtt future. As such it is should be trivial to implement the
ioctl with very little impact.
-chris

-- 
Chris Wilson, Intel Open Source Technology Centre